Melotte 15

[JR]

Melotte 15 lies in the heart of the Heart Nebula IC 1508, some 7,500
light years way.  It's a cluster of newborn stars about 1.5 million
years old within the local gas clouds that have been shaped by the
particle flow and radiation from the stars.

Images were taken last Saturday night with a rare lasting clear sky from
about 9pm, curtailed only by the house getting in the way.

Mono and colour versions uncropped and cropped from the originals are
attached.  Take your pick.  I get most from mono.  Anything purple is my
colour vision!

Esprit 80ed, f5/400, ZWO asi533MC pro OSC, Optolong dual band Ha-Oiii
filter.  26 frames were taken at -10C for 180 seconds each at gain 150. 
The results were calibrated with matching darks and stacked and
processed in Pixinsight (Background neutralisation, background
extraction, colour calibration, channel extraction, per channel
histogram equalisation, the Oiii blue & green channels merged with 60/40
weight for 'blue', channel re-combination with r as luminance, r as
Ha/red, 'blue' as green and green as blue (following the 'Lukomatico'
methodology for Hubble type results from two channel OSC), Pixinsight
script EZ denoise.  Stars were removed to facilitate further processing
using Starnet, followed by curves transformations with Range Masks for
saturation adjustments).

James


--
This email has been checked for viruses by AVG.
https://www.avg.com

[Trev S]

James,

It looks like you are really getting to grips with Pixinsight, you have some very good results.

I particularly like the "starscrop" one, it shows a lot of detail, though it looks like the stars have saturated which is so easy to do and difficult to recover from.

[tcos...@gmail.com]

Hi James
That's a great set of images you have posted! I agree with Trev that you seem to be really getting to grips with Pixinsight. Its also interesting to see the results of Starnet++ for removing stars. Its done a good job on your Heart Nebula image. I downloaded this tool a few weeks ago but have only tried it on a test basis so far. I wonder how it will cope with a field with lots of small stars? I guess also recombining it with the original 'with star' image in a natural way could be a challenge. Separating stars from background and processing the 2 elements separately now seems to be more popular. Also, people seem to be adding LRGB star data to standard narrowband images to make the stars more realistic/normal. All very interesting to see but finding the time to experiment with these methods is difficult.
KR
Tim C

--
You received this message because you are subscribed to the Google Groups "croydonastro" group.
To unsubscribe from this group and stop receiving emails from it, send an email to croydonastro...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/croydonastro/1c962e46-a6f1-f6ea-ad4b-05e2fe87c70a%40aol.com.

[JR]

Thanks very much Tim.

I'm getting there with Pixinsight but there are many parts of my map that are blank outlines.

Starnet does a very good job with most images but has to be used on non-linear images, ie ones with a permanent stretch, not just a screen viewability one. The only minor problem I've had sometimes is for large stars, where a scar from the removal and interpolation may be evident. The routine uses a clone type function to infill using the surrounding local background. It's easy enough to use a clone tool manually to improve the result if needed.

Putting things back together is very simple indeed in Pixinsight using its built in Pixel Maths process. All images in Pixinsight have a name starting with whatever it was first named or changed simply by altering the image identifier eg to 'no_stars' (Pixinsight uses underscores not spaces). Once you have the starless version you subtract it from the version with stars, in a mathematical expression citing the image names eg

with_stars - no_stars

The result is just the stars, which can be named 'stars'.

The no_stars version can then be adjusted quite robustly without affecting the stars, and once you have what is a final version, Pixel Math does the recombination with an equally simple mathematical expression eg

adjnostars + stars

Job done, though you could apply a routine to reduce the number of stars if you want.

Pixel Maths is very powerful. It operates on normalised images, where the values in an image run from 0 to 1. I'm not sure how many shades you get but Pixinsight processing is 32 bits and I think 2^32 is a 10 digit decimal number. The simplest expression is just a numerical value between 0 and 1. For instance 0 gets you a white frame and 0.4 gets you a 'grey' frame. It can get very complex if one is up to that sort of thing.

Just to prove I'm not a fanboy for Pixinsight, I still find Photoshop easier for solar imaging, where parts of images have to be defined and adjusted in a way that Pixinsight masks can't manage (or I can't see how!)

James



> On 5 Feb 2022, at 09:01, tcos...@gmail.com wrote:
>
> Hi James

> To view this discussion on the web visit https://groups.google.com/d/msgid/croydonastro/00e601d81a6e%24e40c0f70%24ac242e50%24%40gmail.com.

[William Bottaci]

Hello James, a very nice image.
I particularly like the detail in the nebula so it looks like you did a good job with processing that, without the stars getting in the way.
One thing though, I notice that none of the stars seem to have any colour; is that the case or am I missing something?
However, the size of the stars is very much controlled, nice and tiny, which must be quite an achievement in itself.

Tim, I recall your 'IC1805 Melotte 15 Heart Nebula' image from 28th September last year - interesting to compare both images - each has something the other doesn’t, which can only be a good thing taking both together
Thanks for sharing James.
William

On Sun, 6 Feb 2022 at 09:29, 'JR' via croydonastro <croydo...@googlegroups.com> wrote:
Thanks very much Tim.
I'm getting there with Pixinsight but there are many parts of my map that are blank outlines.

Starnet does a very good job with most images but has to be used on non-linear images, ie ones with a permanent stretch, not just a screen viewability one.  The only minor problem I've had sometimes is for large stars, where a scar from the removal and interpolation may be evident.  The routine uses a clone type function to infill using the surrounding local background.  It's easy enough to use a clone tool manually to improve the result if needed.

Putting things back together is very simple indeed in Pixinsight using its built in Pixel Maths process.  All images in Pixinsight have a name starting with whatever it was first named or changed simply by altering the image identifier eg  to 'no_stars'  (Pixinsight uses underscores not spaces).  Once you have the starless version you subtract it from the version with stars, in a mathematical expression citing the image names eg

with_stars - no_stars

The result is just the stars, which can be named 'stars'.

The no_stars version can then be adjusted quite robustly without affecting the stars, and once you have what is a final version, Pixel Math does the recombination with an equally simple mathematical expression eg
adjnostars + stars

Job done, though you could apply a routine to reduce the number of stars if you want.

Pixel Maths is very powerful.  It operates on normalised images, where the values in an image run from 0 to 1.  I'm not sure how many shades you get but Pixinsight processing is 32 bits and I think 2^32 is a 10 digit decimal number.  The simplest expression is just a numerical value between 0 and 1.  For instance 0 gets you a white frame and 0.4 gets you a 'grey' frame.  It can get very complex if one is up to that sort of thing.

Just to prove I'm not a fanboy for Pixinsight, I still find Photoshop easier for solar imaging, where parts of images have to be defined and adjusted in a way that Pixinsight masks can't manage (or I can't see how!)
James

On Sat, 5 Feb 2022 at 09:01, <tcos...@gmail.com> wrote:
Hi James
That's a great set of images you have posted! I agree with Trev that you seem to be really getting to grips with Pixinsight. Its also interesting to see the results of Starnet++ for removing stars. Its done a good job on your Heart Nebula image. I downloaded this tool a few weeks ago but have only tried it on a test basis so far. I wonder how it will cope with a field with lots of small stars? I guess also recombining it with the original 'with star' image in a natural way could be a challenge. Separating stars from background and processing the 2 elements separately now seems to be more popular. Also, people seem to be adding LRGB star data to standard narrowband images to make the stars more realistic/normal. All very interesting to see but finding the time to experiment with these methods is difficult.
KR
Tim C

On Wed, 2 Feb 2022 at 16:53, 'JR' via croydonastro <croydo...@googlegroups.com> wrote:
Melotte 15 lies in the heart of the Heart Nebula IC 1508, some 7,500 light years away. It's a cluster of newborn stars about 1.5 million years old within the local gas clouds that have been shaped by the particle flow and radiation from the stars.

[tcos...@gmail.com]

Hi William

I agree that its interesting to compare different images of the same object because there will always be something new and different in each of the images. Re the star colour issue you noted, in processing narrowband images (certainly for Hubble palette colours) star colour has to be adjusted to correct for an over-powering magenta hue which results from combining the 3 narrowband channels. This results in narrowband stars appearing colourless. To remedy this a recent trend has been for narrowband and broadband images to be taken of the same object – narrowband processed as normal and broadband (RGB) channels used solely to give the stars a ‘natural’ broadband colour. I have not tried processing an image using this technique yet, but I have some data I took for the dolphin head nebula to process which has both narrowband and broadband data for this purpose. In theory, having stripped out the stars and processed the nebula in ‘starless mode’ it should be relatively straightforward to add the broadband stars in. However, stars rendered in narrowband usually appear smaller than broadband ones, so some experimentation may be necessary in the combination process.

KR

Tim C

 

From: croydo...@googlegroups.com <croydo...@googlegroups.com> On Behalf Of William Bottaci

Sent: 14 June 2022 17:20
To: croydo...@googlegroups.com

--

You received this message because you are subscribed to the Google Groups "croydonastro" group.
To unsubscribe from this group and stop receiving emails from it, send an email to croydonastro...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/croydonastro/CAHfnR27-WdOG%2BmGqcp15uGxXF6dy_ZQgt_bbxPGFjzz%3D5saS5A%40mail.gmail.com.

[JR]

Hi William

Thanks very much for your kind comments.

You are quite right about star colour.  Tim has summed it up very well.  Though I use a colour camera, it's with the one shot Optolong L-extreme narrowband bi colour filter (combined 7nm Ha and Oiii).

Adding in broadband rgb is the way to go.  It's something I'm investigating as pixinsight can do this (and I have found a tutorial on doing it).  As well I have to overcome the practical constraint of getting the data.  I usually end up taking fewer narrowband images than I would like, as objects get screened by the house or trees, if they are visible at all, and of course sometimes by clouds.  RGB would need another night and there is rarely a run of consecutive good ones, indeed large gaps are the norm.

regards

James

Sent from my iPad

On 14 Jun 2022, at 19:23, tcos...@gmail.com wrote:



To view this discussion on the web visit https://groups.google.com/d/msgid/croydonastro/2133501d8801b%24d8722060%2489566120%24%40gmail.com.

[JR]

Hi William

Just by coincidence I received a link to a talk by Russell Croman yesterday (Gradient Xterminator and now Star and Noise-Xterminator), about the process of removing stars from images.  You mentioned Starnet and being able to process the starless image without affecting the stars.

The talk explained how star removal works using a neural network approach.  This boils down to using pairs of images, original with stars and the same image without stars.  The without stars image is manually prepared, eg using clone stamp etc to remove the star and then fill in the hole.  

The pixels of the original image are 'convolved' to find the best way to transform the overall matrix of pixel values in the original version to the required starless version using Graphic Processing Units.  This is called training the algorithm.  Google has pioneered a lot of the hardware and software for its own marketing purposes, some public access and the latest not.

GPUs are now capable of carrying out multiplications at a rate of at least 3 or 4 trillion a second, in the case of Pixinsight, on pixel values normalised to take a value between 0 and 1.  If you want to halve the luminance pixel value of pure white, 1, you multiply by 0.5 for instance.

The required algorithm is derived by using a Kernel, another smaller matrix, which multiplies the original pixels until the hand derived final starless version is achieved.  The starting Kernel uses random numbers.  Each iteration varies the value of the Kernel and sees if a multi factor outcome measure, in terms of the objective of fewer stars and any other criteria required, is made better or worse. 

The best result is achieved after several days of processing with GPUs.  The process is refined so that it works well enough on a variety of different images - Star Xterminator used about 20 I think.  If problems arise with a class of images, an updated algorithm can be generated.  

The bottom line is:

*  It is brute force AI.  As far as I could tell there is no attempt to derive rules which suggest how to vary the kernel across iterations - more like the infinite number of monkeys typing randomly on their infinite number of typewriters until they write the Complete Works of Shakespeare.  Trillions of iterations are tried until one improves the outcome measure and continues to do so for subsequent iterations - but not a trivial problem to be sure that an optimum is a global optimum rather than a local one.  I expect Google is not unaware of the issue of adding predictive intelligence to the process.  

*  It's impossible to ever know the details of how the starless result was achieved.  The process works and that's as far as the human mind is privy to that of the artificial intelligence.  Perhaps one day, another AI computer, or a whole group of AIs, will be able to understand every step of the whole process and figure out how to work together as a team without being told to or how to.   Then things get scary if you are into the 'singularity', the precise microsecond when Artificial Intelligence achieves artificial self awareness.

*  Star removal AI destroys the scientific value of the original pixels, important if spectroscopy or astrometry is required in scientific applications.  Filling in the holes left by stars can be context sensitive but never on a fully knowable basis.  

* a neural network process is used to make (the word is used advisedly) the recent images of black holes, by convolving the actual data into the form expected by theory - rather circular to my mind.

In the mean time, if Starnet works we should be pleased.  Or maybe not.

James

Sent from my iPad

On 15 Jun 2022, at 07:13, JR <drja...@aol.com> wrote:

Hi William